Three numbers, not one
Pilot is enough to test whether the approach works. Workable is where most projects start. Strong is where more of the same data stops paying off.
Free tool · fine-tuning calculator
How much training data do I need?
Pick what you are fine-tuning and get a realistic range (pilot, workable and strong), with the published guidance behind it where there is some and a clear flag where there isn’t, plus the evaluation set to hold out and what to do next.
- 13 tasks: LLMs, text and vision
- Every range sourced or flagged
- No sign-up, runs in your browser
01 · How it works
Ranges, not a single number
Published figures for fine-tuning data span three orders of magnitude, and both ends are defensible. The calculator shows where your task sits and says how sure that is.
Sourced, or flagged
A green badge means the range comes from documentation or a paper. Amber means no one publishes a number for that task, so the range is a rule of thumb from practice.
Then let your curve decide
Train on 25%, 50% and 100% of your data and plot a held-out score. Still climbing: collect more. Flat: improve quality or variety. That test beats any published figure.
02 · The ranges
Every range in the calculator
Before adjustments for model size or how subtle the classes are. Per-class and per-task figures multiply by the count you enter.
| Task | Pilot | Workable | Strong | Unit | Basis |
|---|---|---|---|---|---|
| Teach one output format or schema | 100 | 300 | 1,000 | examples | Published + rule of thumb |
| One domain task done well (SFT) | 100 | 1,000 | 3,000 | examples | Published + rule of thumb |
| Several related tasks (SFT) | 100 | 1,000 | 3,000 | examples per task | Published + rule of thumb |
| A general-purpose assistant (SFT) | 10,000 | 100,000 | 940,000 | examples | Published + rule of thumb |
| Preference tuning (DPO, ORPO, KTO) | 1,000 | 10,000 | 60,000 | preference pairs | Rule of thumb |
| RL with verifiable rewards (GRPO) | 500 | 5,000 | 10,000 | prompts with a verifier per capability | Published + rule of thumb |
| Teach new domain knowledge (continued pretraining) | 1B | 10B | 100B | tokens | Published + rule of thumb |
| Text classification (BERT-style encoder) | 50 | 200 | 1,000 | labelled examples per class | Rule of thumb |
| Named-entity recognition | 50 | 300 | 1,000 | annotated entity mentions per entity type | Rule of thumb |
| Object detection (YOLO-style) | 100 | 400 | 10,000 | labelled object instances per class | Published + rule of thumb |
| Image classification (fine-tuning a pretrained model) | 30 | 200 | 1,000 | labelled images per class | Rule of thumb |
| Image generation: a subject or a style (LoRA, DreamBooth): one subject or object | 3 | 5 | 20 | images | Published + rule of thumb |
| Image generation: a subject or a style (LoRA, DreamBooth): a consistent style | 20 | 60 | 100 | images | Rule of thumb |
| Image generation: a subject or a style (LoRA, DreamBooth): a robust style, or several concepts | 100 | 400 | 1,000 | images | Rule of thumb |
| Image–text retrieval (fine-tuning CLIP) | 2,000 | 20,000 | 100,000 | image–text pairs | Rule of thumb |
Sources:
- Unsloth: datasets guide
- Unsloth: LoRA hyperparameters guide
- Unsloth: reinforcement learning guide
- Unsloth: continued pretraining
- LIMA: Less Is More for Alignment (arXiv:2305.11206)
- Tülu 3 (arXiv:2411.15124)
- InstructGPT (arXiv:2203.02155)
- UltraFeedback
- OLMo 2 (arXiv:2501.00656)
- Ultralytics: tips for best training results
- DreamBooth (arXiv:2208.12242)
- Diffusers: DreamBooth training
03 · FAQ
Questions about dataset size
How many examples do I need to fine-tune an LLM?
For one narrow behaviour, hundreds to a few thousand. Unsloth recommends at least 100 rows, with over 1,000 preferable, and LIMA reached strong results with exactly 1,000 curated examples on a 65B model. A general-purpose assistant is a different job: Tülu 3’s SFT mixture is 939,344 samples. The deciding question is how many distinct things the model must do.
Does model size change how much data I need?
Generally, larger models pick up a behaviour from fewer examples because they have more latent capability to select from; smaller models benefit more from extra examples and are less forgiving of inconsistency. The calculator applies this as a rule of thumb.
Is more data always better?
No. Every example is a demonstration, so contradictory or sloppy examples teach the model to sometimes do the wrong thing. Having a person read and fix a thousand examples usually beats collecting ten thousand more.
How big should my evaluation set be?
A few hundred hand-checked items you never train on is a reasonable target: enough that differences are not noise, small enough to read every one. For classifiers, make sure every class is represented.
What if I don't have enough data?
Check whether an existing dataset already covers part of the task (search Dataset Finder’s curated catalogue), then order only the gap. The annotation cost estimator shows what that costs in credits.
04 · Keep reading
The guides behind the numbers
More free tools
Other tools for AI training data
Annotation cost estimatorCredits, euros and plan discounts, compared with labelling in-house.EU AI Act dataset documentationA datasheet mapped to Article 10 and Annex IV, downloadable as Markdown or PDF.Annotation guidelines generatorLabelling guidelines with definitions, edge cases and format, exported to Word or PDF.Cohen’s kappa calculatorInter-annotator agreement and weighted kappa, with a confidence interval.